docs(adr): record the memory cost that pays for the disk win

Fingerprinting a tree means holding it, so peak memory rises from one
listing page to every object that survives the filters.

Author: mbevc1

docs/adr/20260911-s3-fingerprint-from-virtual-tree.md

@@ -62,6 +62,6 @@ The attest side is unchanged: `kosli attest artifact --artifact-type dir` still
 - `localPathForS3Key`, `filepath.IsLocal`, the `O_EXCL` open, the `ENOTDIR` branch, `containsSingleFile` and the platform-conditional tests from #1155 are deleted. No code in the S3 path branches on the operating system any more. The codebase does not get smaller, though: the virtual tree, the key rule with its collision reporting, and above all the faithful simulation of `filepathx.Glob` add several hundred lines, most of them owed to reproducing `.kosli_ignore` semantics exactly. That is the price of the compatibility contract, paid once and shared with #1069.
 - The bucket's ignore file is recognised by the exact key `.kosli_ignore`. On disk, `ignoreFilePathInTree` also accepted a case-folded spelling such as `.KOSLI_IGNORE` where the operator's filesystem folded case, so on macOS or Windows such a bucket had its rules applied; now it does not, its exclusions stop applying and its fingerprint moves to the Linux value. Same correction as the key case above, and release-noted with it.
 - `...`, `.. `, `CON`, colon and backslash keys snapshot again. `..` segments and colliding keys remain errors and now name every key involved.
-- Excluded and ignored objects are not downloaded. Each object's temp file is removed once hashed, so peak temp disk falls from the whole bucket to one object, and to the configured budget once #1167 lands.
+- Excluded and ignored objects are not downloaded. Each object's temp file is removed once hashed, so peak temp disk falls from the whole bucket to one object, and to the configured budget once #1167 lands. The other half of that trade is memory: the old loop streamed the listing a page at a time, while a tree cannot be fingerprinted without holding it, so peak memory rises from one listing page to every object that survives the filters, at a few hundred bytes per object on top of its key across the listing, the key-to-path maps, the manifest and the tree. A million objects is a few hundred megabytes. #1167 sizes its budget against this figure.
 - #1069 rebases onto the shared layer: metadata mode becomes a sha256 source plugged into the same list, normalise, exclude, tree pipeline, its key rule disappears in favour of rule 2, and its rejection of buckets with a root `.kosli_ignore` becomes a download of that one object.
 - Delivery is two pull requests. The first carries this decision with sequential downloads, so the security-relevant review is not mixed with performance work. The second, #1167, adds parallel downloads, the byte budget and the flags that tune them.